MCP Connectors: Claude connected to everything

What is MCP?

The Model Context Protocol (MCP) is an open protocol created by Anthropic in late 2024 and adopted today across the industry — including OpenAI and Google. It defines a standard way for an AI model to discover and use external tools: reading email, querying databases, creating tasks, posting to Slack…

Before MCP, every integration was a custom build. Now, any service can publish an MCP server and any compatible AI can use it instantly. There are over a thousand servers available, and the number grows every week.

How to add a connector

In the Claude app, go to Settings → Connectors and pick from the directory. In Claude Code, one command is enough:

The architecture in 20 seconds

An MCP server exposes three things: tools (actions the AI can execute), resources (data it can read) and prompts (reusable templates). The client — Claude, your IDE, or any compatible app — connects via stdio or HTTP and negotiates the available capabilities. Everything with explicit permissions: you decide what the AI can touch.

Security tips

• Only install connectors from trusted sources (official directory, verified repositories).
• Review permissions: a read-only connector can't modify anything.
• For sensitive data, prefer local MCP servers over remote ones.
• In teams, admins can control which connectors are allowed.

What actually changes when you connect Claude to your data

The difference between an AI with connectors and one without isn't convenience, it's category. Without connectors, every conversation starts from zero: you are the channel through which information arrives, so you copy, paste, describe what's in a document and translate into text what already exists in a database. With connectors, the AI goes and gets it. It stops being someone you brief and becomes something that operates on your systems.

That jump has an immediate practical consequence: the shuffling tasks disappear. "Look at this week's emails from this client, cross-reference what they asked for against their order status and tell me what's outstanding" used to be half an hour of tab-hopping. With Gmail and the database connected, it's one sentence. Not because the model got smarter, but because it finally has access to the data it's supposed to reason about.

What a connector can do, and what it can't

It's worth understanding how responsibility is divided. An MCP server declares which operations exist and with what parameters; the model decides when to use them; and the client — the Claude app, your IDE — executes and returns the result. The model never touches your systems directly: it asks, and something else runs. That separation is what makes the design auditable, because every action passes through a point where it can be logged or blocked.

The important limit is that a connector can only do what the server exposes. If the Gmail server includes no operation for deleting mail, no prompt will get the AI to delete any. That's a virtue, not a gap: when evaluating a connector, the first thing to look at isn't what it promises but which operations it declares and what permissions it asks for in order to run them.

Permissions: the part to read slowly

A connector is authorised with your account's permissions, so it inherits your capacity to do damage. The rule is the same as for any third-party integration, but with an AI in the loop it's worth being stricter than usual.

Always start with read-only scopes and widen them once you've confirmed the workflow works. Many connectors let you pick the scope at the moment of authorising. Be wary of unofficial servers asking for full access to an account holding sensitive data: installing an MCP server means running third-party code on your machine, and it deserves the same scrutiny you'd give any dependency. Review periodically what you have connected and remove whatever you tried once and never used again. And in a work environment, remember the connector sees what your user sees: if your account can reach organisation-wide shared folders, so can the connector.

The counterweight is that consequential actions — sending, publishing, deleting, paying — should require explicit confirmation before they run. A good client asks you. If one doesn't ask, that tells you something about the client.

Where to start if you've never installed one

The classic mistake is connecting everything on day one and then never using any of it. The opposite works far better: pick a single connector, the one matching the application you spend most time in, and use it for a whole week until the novelty turns into habit. Calendar and document connectors tend to be the best first step because the benefit shows on day one and the risk is low.

When that first one starts feeling limited, add the second thinking about the combination rather than the app: the real value appears when two connectors intersect — mail and tasks, repository and documentation, store and spreadsheet — because that's exactly where the purely manual work used to be.

And if what you want is for the AI to always do the same thing with that connected data, the natural next step is skills: connectors give it access, skills give it judgement. The guide to creating your own skills explains how to combine them, and the glossary clears up the protocol terminology if any term trips you up.